Investigating learning rates for evolution and temporal difference learning

Evidently, any learning algorithm can only learn on the basis of the information given to it. This paper presents a first attempt to place an upper bound on the information rates attainable with standard co-evolution and with TDL. The upper bound for TDL is shown to be much higher than for co-evolution. Under commonly used settings for learning to play Othello for example, TDL may have an upper bound that is hundreds or even thousands of times higher than that of coevolution. To test how well these bounds correlate with actual learning rates, a simple two-player game called Treasure Hunt is developed. While the upper bounds cannot be used to predict the number of games required to learn the optimal policy, they do correctly predict the rank order of the number of games required by each algorithm.